Abstract
Most of the computer vision applications for human activity recognition exploit the fact that body features calculated from a 3D skeleton increase robustness across persons and can lead to higher performance. However, their success in activity recognition, including falls, depends on the correspondence between the human activities and the used joint/part features. To provide for this correspondence, we experimentally evaluate in this paper skeleton features-based fall detection by comparing fall detection performance for different combinations of skeleton features used in previous related works. We determine the skeleton features that best distinguish fall from non-fall frames, and the best performing classifier. In this endeavor, we followed the classical five steps of supervised machine learning: (1) we collected a learning data composed of 42 fall and 37 non-fall videos from FallFree; (2) we extracted and (3) preprocessed the skeleton data of the training set; (4) we extracted each possible skeleton feature; finally (5) we evaluated all extracted and selected features using two main experiments; one of them based on neighborhood component analysis (NCA). In this evaluation, we show that fall detection based on skeleton features has very encouraging accuracy that varies depending on the used features. More specifically, we recommend the following features: 12 features that resulted from NCA experiment, original and normalized distance from Kinect, and the seven features of the upper body part. These features ranked 1st, 2nd, 4th, and 8th on 22 feature sets, with accuracies 99.5%, 99.4%, 97.8%, and 94.5%, respectively. In addition, random forest is the best performing classifier.